Structural damage classification identification method and system based on deep learning classification algorithm

Abstract

The present invention discloses a structural damage grading identification method and system based on a deep learning classification algorithm, which relates to the field of structural health monitoring technology. The structural damage grading identification method includes the following steps: pre-processing and enhancing the structural damage image data in the primary data set to form a secondary data set; retaining the best quality structural damage image as the tertiary data set; taking the category with the largest classification probability value as the damage type of the damaged image block; splicing the undamaged image block with the marked damaged image block to obtain the final identification result. The present invention implements a multi-level structural damage identification function based on a deep learning classification algorithm. Compared with traditional single-layer models or other existing detection models, its deep neural network model has the characteristics of gradually refined detection and identification through modular design, and has the ability to automatically identify structural damage adaptively with various types, locations and degrees of structural damage.

Description

Technical Field

[0001] The present invention relates to the technical field of structural health monitoring, and in particular to a structural damage classification identification method and system based on a deep learning classification algorithm. Background Art

[0002] With the continuous expansion and increasing complexity of modern engineering structures, structural health monitoring has become a key means of ensuring structural safety and extending their service life. Demand for structural health monitoring systems is increasing in fields such as civil engineering, aerospace, and marine engineering. Traditional structural damage identification techniques primarily rely on manual inspection, analysis based on vibration signals, or simple image processing methods. While these methods can identify structural damage to a certain extent, they often suffer from shortcomings such as low recognition efficiency, long detection cycles, insufficient ability to distinguish between damage types, and susceptibility to environmental noise.

[0003] Manual inspection usually requires professionals to use portable equipment or visual inspection, which is not only time-consuming and labor-intensive, but also the inspection results largely rely on the experience and subjective judgment of the inspectors, making it difficult to ensure the comprehensiveness and consistency of the inspection. Structural damage detection methods based on vibration signals usually judge the health status of the structure by analyzing the vibration response of the structure under external forces. However, vibration signals are easily interfered by environmental noise and have limited ability to identify minor or complex damage. In addition, traditional image processing methods mainly rely on low-level image processing technologies such as edge detection and image segmentation. These methods often perform poorly in complex backgrounds and have difficulty accurately identifying various types of structural damage.

[0004] In recent years, with the development of artificial intelligence (AI) technology, deep learning has demonstrated powerful image recognition and classification capabilities in computer vision. This has brought new development opportunities for image recognition technology in structural health monitoring. In particular, deep learning can automatically extract the characteristics of structural damage by learning from a large number of annotated samples, enabling accurate identification of damage type, location, and severity. Compared with traditional methods, deep learning-based damage identification technology not only significantly improves recognition accuracy and robustness, but also adapts to complex and changing environmental conditions.

[0005] However, the application of deep learning technology also faces several challenges. First, the performance of deep learning models is highly dependent on large-scale, high-quality training datasets, while data on structural damage in real-world projects is often scarce and unbalanced. Second, training deep learning models requires significant computing resources, and the complexity of the models may result in insufficient real-time performance. Furthermore, maintaining the long-term effectiveness of the models as the structural environment changes is a pressing issue.

[0006] Currently, no effective solutions have been proposed for the problems in related technologies. Summary of the Invention

[0007] In response to the problems in the related art, the present invention proposes a structural damage classification identification method and system based on a deep learning classification algorithm to overcome the above-mentioned technical problems existing in the existing related art.

[0008] To this end, the specific technical solutions adopted in the present invention are as follows:

[0009] According to one aspect of the present invention, a structural damage classification identification method based on a deep learning classification algorithm is provided, the structural damage classification identification method comprising the following steps:

[0010] S1. Acquire a structural damage image dataset as a primary dataset, and preprocess and enhance the structural damage image data in the primary dataset to form a secondary dataset;

[0011] S2. Use the screening technology integrated with Laplace transform algorithm to preliminarily screen the secondary data set and retain the best quality structural damage images as the tertiary data set;

[0012] S3. Based on the deep learning classification algorithm and the three-level dataset, a primary recognition model and a secondary recognition sub-model are constructed; the primary recognition model is used to perform preliminary recognition and classification on the three-level dataset to obtain undamaged image blocks and damaged image blocks, and the damaged image blocks are input into the secondary recognition sub-model for secondary recognition and classification;

[0013] S4. Based on the deep convolutional neural network, feature extraction and classification are performed on the results of the secondary recognition and classification. The category with the largest classification probability value is used as the damage type of the damaged image block, and the damaged image block is labeled; the undamaged image block is spliced ​​with the labeled damaged image block to obtain the final recognition result.

[0014] Furthermore, obtaining a structural damage image dataset as a primary dataset, and preprocessing and enhancing the structural damage image data in the primary dataset to form a secondary dataset includes the following steps:

[0015] S11. Obtain a structural damage image dataset as a primary dataset and set up a generative adversarial network;

[0016] Among them, the generative adversarial network includes a generator and a discriminator;

[0017] S12, inputting the primary dataset as a training set into the generative adversarial network for training, using the generator to generate structural damage image data samples, which are then distinguished by the discriminator;

[0018] S13. After the training is completed, use the generator to generate new structural damage image data samples, input random noise, add the new structural damage image data samples to the original primary dataset, and adjust the labels to form a secondary dataset.

[0019] Furthermore, the secondary dataset is preliminarily screened using the screening technology integrated with the Laplace transform algorithm, and the best quality structural damage images are retained as the tertiary dataset, which includes the following steps:

[0020] S21. Using a Laplace transform algorithm to mark structural damage image data whose clarity is lower than a preset threshold, and manually screening and confirming the marked structural damage image data, filtering out images unrelated to structural damage, obtaining and retaining the structural damage image with the best quality;

[0021] S22. Store the retained best-quality structural damage images in a separate folder to form the final three-level dataset.

[0022] Furthermore, based on the deep learning classification algorithm and the three-level data set, a first-level recognition model and a second-level recognition sub-model are constructed; the first-level recognition model is used to perform preliminary recognition and classification on the three-level data set to obtain damage-free image blocks and damaged image blocks, and the damaged image blocks are input into the second-level recognition sub-model for secondary recognition and classification, including the following steps:

[0023] S31. Select the ResNet-50 architecture as the basic architecture, label the three-level dataset as damage-free and damage-affected, divide the best-quality structural damage image into several image blocks, and use each image block as the input of the first-level recognition model;

[0024] S32. Use ResNet-50 to process the input image blocks layer by layer to extract high-level structural features;

[0025] S34, use the fully connected layer to output the probability value of the two classifications, and convert the output of the first-level recognition model into a probability distribution through the softmax activation function;

[0026] S35. Use the binary weighted cross entropy loss function to measure the difference between the output of the first-level recognition model and the true label, and update the weight of the first-level recognition model through back propagation and gradient descent method;

[0027] S36, performing binary classification on all image blocks in the third-level data set based on the first-level recognition model to screen out damage-free image blocks and damaged image blocks;

[0028] S37. Based on the damaged image blocks, construct several parallel secondary recognition sub-models, and classify different types of structural damage into several damage levels through each secondary recognition sub-model.

[0029] Furthermore, ResNet-50 is used to process the input image blocks layer by layer to extract high-level structural features including:

[0030] The input image block passes through the convolution layer, pooling layer and residual block of ResNet-50 to extract low-level features layer by layer and gradually aggregate them into high-level structural features.

[0031] Furthermore, the formula for the binary weighted cross entropy loss function is:

[0032]

[0033] Where Loss(1) represents the loss function value of the first-level recognition model;

[0034] N represents the total number of input image blocks for one training;

[0035] y i Represented as the true label of the i-th image block;

[0036] p i It is represented as the predicted category distribution probability of the first-level recognition model of the i-th image block.

[0037] Furthermore, based on the damaged image blocks, a plurality of parallel secondary recognition sub-models are constructed, and different types of structural damage are classified into a plurality of damage levels by each secondary recognition sub-model, including the following steps:

[0038] S371. Based on the damaged image block, construct several parallel secondary recognition sub-models, and select a ResNet-50 architecture for each secondary recognition sub-model;

[0039] S372. Classify and label different types of structural damage, and divide structural damage into several damage levels.

[0040] Furthermore, based on a deep convolutional neural network, feature extraction and classification are performed on the results of the secondary recognition and classification. The category with the largest classification probability value is used as the damage type of the damaged image block, and the damaged image block is labeled. The damage-free image block is spliced ​​with the labeled damaged image block to obtain the final recognition result, which includes the following steps:

[0041] S41, based on the convolution layer of the deep convolutional neural network, a sliding window scan is performed on the input damaged image block, and a number of filters are used to extract the local features of the damaged image block;

[0042] S42, downsampling the feature damaged image block using a maximum pooling layer to reduce the size of the feature damaged image block;

[0043] S43. After multi-layer convolution and pooling, the extracted high-level feature damaged image is flattened into a one-dimensional vector and passed to the fully connected layer. The fully connected layer maps the feature damaged image to the final category space. Based on the Softmax activation function, the output value of the secondary recognition sub-model is converted into the probability distribution of each damage category.

[0044] S44. Use the multi-class weighted cross entropy loss function to measure the difference between the output of the first-level recognition model and the true label, and update the weight of the first-level recognition model through back propagation and gradient descent method;

[0045] S45. Based on the Softmax layer, the probability distribution of each type of damage for each damaged image block is output, the category with the highest probability value is selected as the final damage type of the damaged image block, and corresponding labeling is performed;

[0046] S46: splicing the intact image block with the marked damaged image block to form a result image consistent with the specifications of the original image as the final recognition result.

[0047] Furthermore, the formula of the multi-class weighted cross entropy loss function is:

[0048]

[0049] Where Loss(2) represents the loss function value of each secondary recognition sub-model;

[0050] N represents the total number of input image blocks for one training;

[0051] M represents the total number of categories;

[0052] y i，c Represented as the true label of the i-th image block on the c-th category;

[0053] p i，c It is represented as the category distribution probability predicted by the secondary recognition sub-model of the i-th image block on the c-th category.

[0054] According to another aspect of the present invention, a structural damage classification identification system based on a deep learning classification algorithm is also provided. The structural damage classification identification system includes:

[0055] A data acquisition module is used to acquire a structural damage image data set as a primary data set, and preprocess and enhance the structural damage image data to form a secondary data set;

[0056] A data screening module is used to perform preliminary screening on the secondary data set using a preset screening technology and retain the best quality structural damage images as the tertiary data set;

[0057] The data classification module is used to build a primary recognition model and a secondary recognition sub-model using a deep learning classification algorithm and a three-level data set. The primary recognition model is used to perform preliminary recognition and classification on the three-level data set to obtain intact image blocks and damaged image blocks. The damaged image blocks are then input into the secondary recognition sub-model for secondary recognition and classification.

[0058] The result recognition module is used to extract features and classify the results of secondary recognition and classification based on a deep convolutional neural network. The category with the largest classification probability value is used as the damage type of the damaged image block, and the damaged image block is labeled; the damage-free image block is spliced ​​with the labeled damaged image block to obtain the final recognition result.

[0059] The beneficial effects of the present invention are:

[0060] 1. This invention implements multi-level structural damage identification based on a deep learning classification algorithm. Compared to traditional single-layer models or other existing detection models, its deep neural network model, through a modular design, features progressively refined detection and identification, with the ability to adaptively identify structural damage across a wide range of damage types, locations, and severity levels. This automated and efficient structural damage detection is achieved, enabling automatic identification of damage type and severity, allowing the model to initially perform coarse-grained classification followed by more precise classification at a finer level. This not only reduces computing resource consumption but also improves detection accuracy. The combination of efficient screening and refined classification makes the entire system more modular, flexible, and scalable, with the potential to further expand the scope of identifiable categories and the database through the addition of additional subnetworks.

[0061] 2. This invention achieves efficient and accurate detection of structural surface damage by constructing a primary recognition model and multiple secondary recognition sub-models. This approach offers advantages such as progressively refined detection, modular design, optimized computing resources, and a low false positive rate. While maintaining overall detection efficiency, it significantly improves detection accuracy and robustness, enabling efficient identification of various damage types, locations, and degrees across engineering structures, enhancing both detection precision and efficiency.

[0062] 3. This invention employs hierarchical detection and recognition. First, a primary recognition model rapidly identifies areas of potential damage, reducing interference from irrelevant areas. Furthermore, multiple parallel secondary recognition sub-models further refine the classification, improving the accuracy of identifying different damage types and levels. This hierarchical structure significantly improves overall detection accuracy, particularly in complex and diverse damage scenarios, enabling more accurate distinction of subtle damage types.

[0063] 4. By decomposing the detection process into two stages and combining efficient screening with fine classification, the present invention can efficiently and accurately identify various types of structural damage, and has strong scalability and adaptability. It not only improves the accuracy and efficiency of detection, but also screens out damaged areas through simple and fast binary classification before conducting more detailed detection. It can significantly reduce the computational burden and optimize resource utilization, which will significantly enhance the intelligence and automation level of structural health monitoring and provide important technical support for the safety management of engineering structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0065] Figure 1 is a flowchart of a structural damage classification identification method based on a deep learning classification algorithm according to an embodiment of the present invention;

[0066] Figure 2 is a principle block diagram of a structural damage classification identification system based on a deep learning classification algorithm according to an embodiment of the present invention;

[0067] Figure 3 It is the overall architecture of the primary recognition model and the secondary recognition sub-model according to an embodiment of the present invention;

[0068] Figure 4 is a workflow diagram of a primary recognition model and a secondary recognition sub-model according to an embodiment of the present invention;

[0069] Figure 5 is a schematic diagram illustrating the scalability of the primary recognition model and the secondary recognition sub-model according to an embodiment of the present invention;

[0070] Figure 6 Schematic diagram of the result output of the primary recognition model and the secondary recognition sub-model according to an embodiment of the present invention.

[0071] In the picture:

[0072] 1. Data acquisition module; 2. Data screening module; 3. Data classification module; 4. Result recognition module. DETAILED DESCRIPTION

[0073] To further illustrate each embodiment, the present invention provides drawings, which are part of the disclosure of the present invention. They are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. By referring to these contents, ordinary technicians in this field should be able to understand other possible implementation methods and advantages of the present invention. The components in the figures are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0074] According to an embodiment of the present invention, a method and system for structural damage classification identification based on a deep learning classification algorithm are provided.

[0075] The present invention will now be further described with reference to the accompanying drawings and specific embodiments. Figure 1 As shown, according to the structural damage classification identification method based on the deep learning classification algorithm according to an embodiment of the present invention, the structural damage classification identification method includes the following steps:

[0076] S1. Acquire a structural damage image dataset as a primary dataset, and preprocess and enhance the structural damage image data in the primary dataset to form a secondary dataset;

[0077] It should be explained that the training dataset is prepared according to the task requirements, and a large amount of structural damage image data is collected through various channels, including but not limited to damage types such as cracks, corrosion, material fatigue, deformation, spalling, and connection failure. These datasets come from various engineering structures such as buildings, bridges, and mechanical equipment, covering different material types and damage levels, forming a primary dataset. The image data in the primary dataset is preprocessed and enhanced. Data enhancement refers to the processing or transformation of the existing dataset, including the generation of new data samples through geometric transformations or generative neural networks, to increase data diversity, thereby improving the generalization ability of the machine learning model, and forming a data-enhanced secondary dataset.

[0078] It needs to be explained that in data enhancement, classical data enhancement methods (i.e., geometric transformation) and generative (i.e., generative neural network) data enhancement methods are used to increase data diversity. Classic data enhancement methods include image scaling, cropping, grayscale adjustment, rotation, translation, flipping, color adjustment and other operations, while generative data enhancement methods use deep learning technologies such as generative adversarial networks to generate new structural damage image samples to enrich the representation of the data set.

[0079] It should be explained that geometric transformation mainly transforms existing data (usually image data) by rotating, scaling, translating, shearing, flipping, etc. to generate more data samples. Through these transformations, although the data samples have changed, their basic semantic features are still maintained. The implementation steps include:

[0080] Step 1: Data preparation: Prepare an existing dataset, usually image data, but it can also be other types of data.

[0081] Step 2: Select a geometric transformation method: Select an appropriate geometric transformation method, such as rotation, scaling, translation, flipping, etc. You can use one method alone or combine multiple methods.

[0082] 1) Rotation: Randomly rotate the image within a certain range, such as ±30 degrees.

[0083] 2) Scaling: Randomly scale the image within a certain range, such as enlarging or reducing the image to 0.8 times or 1.2 times the original image.

[0084] 3) Translation: Randomly translate the image, moving it horizontally or vertically.

[0085] 4) Flip: Flip the image horizontally or vertically.

[0086] 5) Shear: Apply a shear transformation to the image.

[0087] Step 3. Apply transformation: Use programming tools (such as Python's opencv library, PIL library, and imgaug library) to transform the data samples one by one to generate new data samples.

[0088] Step 4: Save and expand the data set: Add the newly generated samples to the original data set and label the new data.

[0089] Step 5: Generate new data samples: After training is complete, use the generator to generate new data samples. By inputting random noise, the generator will generate samples that match the distribution of the original data. Add the newly generated data to the original dataset and adjust the labels appropriately.

[0090] S2. Use the screening technology integrated with Laplace transform algorithm to preliminarily screen the secondary data set and retain the best quality structural damage images as the tertiary data set;

[0091] It should be explained that the image data in the secondary dataset are further manually screened to remove data of poor quality or that cannot correctly reflect structural damage generated during the data enhancement process, and only retain image data containing structural damage to ensure the purity and relevance of the dataset, forming a screened tertiary dataset, which is used for training deep learning models (i.e., the primary recognition model and the secondary recognition sub-model), and the tertiary data are divided into a training dataset and a verification dataset for deep learning model training in a ratio of 8:2.

[0092] It should be explained that the image data in the secondary dataset was further manually screened to remove data of poor quality or that could not accurately reflect structural damage generated during the data enhancement process. Only image data containing structural damage was retained to ensure the purity and relevance of the dataset. The filtered tertiary dataset includes:

[0093] Removing low-quality images: During the data augmentation process, some images of poor quality may be generated, such as blurred, distorted, or noisy images, which may not accurately reflect structural damage. The goal of screening is to remove these low-quality samples.

[0094] Ensure accurate reflection of structural damage: retain images that correctly reflect structural damage and remove samples that do not contain damage or have unclear damage to ensure the relevance of the dataset.

[0095] Improve the purity of the dataset: The images finally retained should be highly relevant to the task objective (structural damage detection), avoid interference from irrelevant samples, and improve the overall quality of the tertiary dataset.

[0096] Specific screening steps include:

[0097] Step 1. Preliminary inspection and filtering: First, a comprehensive preliminary inspection of the secondary dataset is performed, and images are filtered using a combination of automated detection and manual screening.

[0098] Clarity detection: Automated tools are used to detect image clarity and remove blurry or poor-quality images. Laplace transforms or other image gradient-based algorithms can be used to automatically flag images with clarity below a preset threshold.

[0099] Step 2: Manual screening: The goal of manual screening is to filter out images that are not related to structural damage through manual observation, and focus on retaining images that reflect damage.

[0100] Manual image inspection: Manually inspect each image, focusing on checking whether there are obvious structural damage features in the image, such as cracks, deformation, and breakage.

[0101] Step 3: Delete images without damage or misclassification: After manual screening, delete images without obvious structural damage. These images may be erroneous samples caused by data augmentation or do not meet the requirements of the damage identification task.

[0102] Step 4: Form a three-level dataset: Save the retained high-quality images as a three-level dataset and store these images in a separate folder.

[0103] S3. Based on the deep learning classification algorithm and the three-level dataset, a primary recognition model and a secondary recognition sub-model are constructed; the primary recognition model is used to perform preliminary recognition and classification on the three-level dataset to obtain undamaged image blocks and damaged image blocks, and the damaged image blocks are input into the secondary recognition sub-model for secondary recognition and classification;

[0104] It should be explained that building a first-level recognition model includes:

[0105] 1) Model Architecture Selection: Select the ResNet-50 architecture. ResNet-50 is a residual network-based model that uses skip connections to address the vanishing gradient problem in deep networks. It effectively extracts deep features from images.

[0106] 2) Data Preparation: Prepare a binary classification dataset, labeled as "undamaged" and "damaged." The dataset can include actual image data and enhance the data diversity by performing image augmentation (e.g., rotation, scaling, and flipping).

[0107] 3) Model input: The input image is divided into several small blocks (image blocks), each of which serves as the input to the model. The input size is generally fixed, such as 224×224 pixels.

[0108] 4) ResNet-50 Hierarchical Structure: The input image passes through ResNet-50's 50 layers of residual blocks, extracting image features layer by layer. As the network deepens, the model is able to extract high-level image features, thereby distinguishing whether there is structural damage in the image.

[0109] 5) Classification Layer and Activation Function: The final layer is a fully connected layer, which outputs probabilities for two categories (no damage or damage). A softmax activation function is used to convert the output into a probability distribution for binary classification.

[0110] 6) Loss function: The cross entropy loss function (i.e., the binary weighted cross entropy loss function) is used to measure the difference between the prediction results of the first-level recognition model and the true label.

[0111] 7) Model optimization: Use backpropagation and gradient descent methods (such as Adam or SGD optimizer) to update the model weights, reduce the loss value, and gradually improve the classification accuracy of the first-level recognition model.

[0112] It should be explained that the damaged image block is input into the secondary recognition sub-model for secondary recognition and classification, including:

[0113] Step 1: Build the sub-model architecture: Each sub-model uses the ResNet-50 architecture, but the final fully connected layer output is different, corresponding to different types of damage and their fine-grained classification.

[0114] For example, the output of the crack identification sub-model is three categories (primary cracks, secondary cracks, and tertiary cracks), and the output of the concrete spalling identification sub-model is also three categories (primary spalling, secondary spalling, and tertiary spalling).

[0115] Step 2: Data Preparation: Preprocess the image blocks identified as "damaged" by the primary recognition model to ensure they are of a consistent size (e.g., 224×224 pixels) and pass them to the corresponding sub-model. The training dataset for each sub-model should include different fine-grained classification annotations for that type of damage. For example, the dataset for the crack recognition model needs to be annotated as primary, secondary, and tertiary cracks.

[0116] Step 3: Train sub-models: Each sub-model is trained on its corresponding damage type dataset, and the model weights are optimized using the cross-entropy loss function and optimizer.

[0117] Step 4: Secondary recognition and prediction: When the primary recognition model determines that a certain area is damaged, the image block will be passed to the corresponding sub-model for secondary recognition.

[0118] For example, when the crack recognition model receives a "damaged" image patch, it will perform fine-grained classification on the image patch and output the crack level (such as first-level crack, second-level crack, or third-level crack).

[0119] Step 5: Parallel processing of multiple sub-models: The sub-models operate in parallel, each responsible for handling different types of damage. For areas identified as "damaged" by the primary recognition model, the image blocks are sent to the corresponding sub-model based on the damage type. Similar fine-grained classification is performed in each sub-model.

[0120] Step 6. Result summary: Finally, the fine-grained classification results of each sub-model are summarized and displayed on the original image as the final result output.

[0121] Specifically, such as Figure 6 As shown in the figure, further segmentation means that after the first-level recognition model identifies "damaged" image blocks, the second-level recognition sub-model performs a more detailed classification of these damaged areas. The segmentation process is to further classify a certain type of damage (such as cracks, concrete spalling, exposed rebar, etc.) into different levels, such as first-level damage, second-level damage, and third-level damage. The purpose of segmentation is to more accurately describe the extent or nature of damage, so that more specific treatment measures can be taken. It includes:

[0122] Step 1: Define segmentation criteria:

[0123] First, clarify the subdivision criteria for each type of injury. For example:

[0124] Crack subdivision standards:

[0125] Level 1 crack: crack width is less than 1 mm.

[0126] Secondary cracks: crack width is between 1-3 mm.

[0127] Level 3 cracks: crack width is greater than 3 mm.

[0128] Concrete spalling subdivision standards:

[0129] First-level spalling: The spalling area is small and the impact is small.

[0130] Secondary spalling: The spalling area is medium and may affect the structural surface.

[0131] Level 3 spalling: The spalling area is large, affecting the structural performance.

[0132] These criteria need to be derived from experience, industry standards, or engineering studies and provide a clear basis for classifying models.

[0133] Step 2: Data annotation:

[0134] Segment the dataset: Prepare damaged images and annotate each image based on the damage criteria. For crack datasets, annotate them as "level 1 crack," "level 2 crack," or "level 3 crack." For concrete spalling, annotate them as different spalling levels.

[0135] Data augmentation: To improve the diversity of the dataset, data augmentation methods (such as rotation, scaling, and flipping) can help generate more damage image samples at different angles and positions, allowing the model to learn more fine-grained features.

[0136] It should be explained that the overall structure of the deep learning classification algorithm is mainly composed of a first-level recognition model and multiple parallel second-level recognition sub-models. Among them, the first-level recognition model processes the input data first through the first-level recognition model. The main task of the first-level recognition model is to perform preliminary analysis and recognition of the input data, and only classify the input data into two categories (no damage or damage), quickly excluding the undamaged areas. For the image blocks judged to be undamaged, no subsequent processing is performed and they are saved in the final result. Only the areas judged to be damaged enter the second-level recognition sub-model.

[0137] Specifically, the specific implementation steps of the secondary recognition model include:

[0138] Step 1: Data preprocessing: For image blocks that are judged as "damaged" from the first-level recognition model, preprocessing is required to adapt to the input of the second-level recognition model.

[0139] Resize image: Resize the input image to a fixed size (e.g. 224×224 pixels).

[0140] Normalize image: normalize the image.

[0141] Step 2: Build a secondary recognition model: The secondary recognition model uses ResNet-50 as the basic architecture and modifies its final output layer according to the specific task to adapt to the fine-grained classification task.

[0142] Step 3. Define the loss function and optimizer: The secondary recognition model uses the cross-entropy loss function to calculate the difference between the model prediction and the actual label. The optimizer uses Adam or SGD to adjust the model weights.

[0143] Step 4: Model training: By training on a dataset labeled with fine-grained categories, the secondary recognition model can gradually learn and accurately classify different degrees of damage.

[0144] Step 5: Model Evaluation and Prediction: After training is complete, the model is used for prediction. The secondary recognition model outputs a fine-grained damage classification result based on the input image blocks that have been identified as "damaged" by the primary recognition model.

[0145] Step 6. Result summary and subsequent processing: After the secondary recognition model outputs the fine-grained damage classification results, the classification results are marked on the original image as the final result output.

[0146] It needs to be explained that, Figure 3 As shown, the first-level recognition model is a binary classification deep learning model that classifies image blocks as either undamaged or damaged. The first-level recognition model uses a binary weighted cross-entropy loss function for convergence calculations. Image blocks judged as undamaged by the first-level recognition model are not processed and the original image is saved until the final result is output. Image blocks judged as damaged by the first-level recognition model are further input into multiple parallel second-level recognition sub-models for refined classification and recognition. The second-level recognition sub-model includes multiple parallel sub-models, each of which performs fine-grained classification for different types of damage. After processing the output of the first-level recognition model, each sub-model performs secondary recognition on areas judged as damaged by the first-level recognition model, determining the likelihood that the damage belongs to the damage category corresponding to each sub-model.

[0147] For example, Figure 4As shown: Secondary identification sub-model #1 corresponds to crack category: it processes the first category of damage, further subdividing it into primary cracks, secondary cracks, and tertiary cracks, and outputs classification result #1.

[0148] Secondary identification sub-model #2 corresponds to the concrete spalling category: it processes the second category of damage, further subdividing it into primary spalling, secondary spalling, and tertiary spalling, and outputs classification result #2.

[0149] Secondary identification sub-model #3 corresponds to the exposed rebar category: it processes the third category of damage, further subdividing it into first-level exposed rebar, second-level exposed rebar, and third-level exposed rebar, and outputs classification result #3.

[0150] Secondary identification sub-model #4 corresponds to the steel bar corrosion category: it processes the fourth category of damage and further subdivides it into first-level steel bar corrosion, second-level steel bar corrosion, and third-level steel bar corrosion, and outputs classification result #4.

[0151] Secondary identification sub-model #5 corresponds to the component concrete crushing category: it processes the fifth category of damage and further subdivides it into primary concrete crushing, secondary concrete crushing, and tertiary concrete crushing, and outputs classification result #5.

[0152] Secondary identification sub-model #6 corresponds to the obvious deformation category of the component: it processes the damage of the sixth category and further subdivides it into primary deformation, secondary deformation, and tertiary deformation, and outputs classification result #6. ...

[0154] Secondary identification sub-model #n corresponds to the category of obvious deformation of the component: it processes the damage of the nth category, further subdivides it into primary damage, secondary damage and tertiary damage, and outputs classification result #n.

[0155] It should be explained that although this embodiment only lists the above categories #1 to #n, similar category divisions and category supplements (i.e., new categories) based on this concept, such as Figure 5 As shown, for example:

[0156] Secondary identification sub-model #n+1 corresponds to the category of obvious deformation of the component: it processes the damage of the n+1th category, further subdivides it into primary damage, secondary damage, and tertiary damage, and outputs classification result #n+1.

[0157] Secondary identification sub-model #n+2 corresponds to the category of obvious deformation of the component: it processes the damage of the n+2th category, further subdivides it into primary damage, secondary damage, and tertiary damage, and outputs classification result #n+2. ...

[0159] Secondary identification sub-model #n+m corresponds to the category of obvious deformation of the component: it processes the damage of the n+mth category, further subdivides it into primary damage, secondary damage, and tertiary damage, and outputs the classification result #n+m.

[0160] It's important to clarify that a binary classification deep learning model refers to a deep learning algorithm used to perform a binary classification on input data. Specifically, it classifies an input image into one of two categories: "undamaged" or "damaged." This binary classification deep learning model uses a deep neural network to automatically extract image features. Using a multi-layer neural network, the original image is transformed into a more abstract feature representation. Finally, a fully connected layer and classifier are used to output the probability of each category. The category with the highest probability is the classification predicted by the binary classification deep learning model.

[0161] S4. Based on the deep convolutional neural network, feature extraction and classification are performed on the results of the secondary recognition and classification. The category with the largest classification probability value is used as the damage type of the damaged image block, and the damaged image block is labeled; the undamaged image block is spliced ​​with the labeled damaged image block to obtain the final recognition result.

[0162] It should be explained that each secondary recognition sub-model uses a deep convolutional neural network to extract features and classify the input image blocks, outputting probability values ​​for each category. Ultimately, the category with the largest classification probability value is used as the damage type for the image block and annotated on the image block. Furthermore, each secondary recognition sub-model uses a multi-classification weighted cross-entropy loss function for convergence calculations. Finally, all image blocks judged as undamaged by the primary recognition model, as well as damaged image blocks labeled with the damage types output by each secondary recognition sub-model, are concatenated to form a result image that is the same size as the original image. This is the overall recognition output result of the secondary recognition sub-model.

[0163] Preferably, obtaining a structural damage image dataset as a primary dataset, and preprocessing and enhancing the structural damage image data in the primary dataset to form a secondary dataset comprises the following steps:

[0164] S11. Obtain a structural damage image dataset as a primary dataset and set up a generative adversarial network;

[0165] Among them, the generative adversarial network includes a generator and a discriminator;

[0166] S12, inputting the primary dataset as a training set into the generative adversarial network for training, using the generator to generate structural damage image data samples, which are then distinguished by the discriminator;

[0167] S13. After the training is completed, use the generator to generate new structural damage image data samples, input random noise, add the new structural damage image data samples to the original primary dataset, and adjust the labels to form a secondary dataset.

[0168] Preferably, using the screening technology integrated with the Laplace transform algorithm to preliminarily screen the secondary data set and retaining the best quality structural damage images as the tertiary data set includes the following steps:

[0169] S21. Using a Laplace transform algorithm to mark structural damage image data whose clarity is lower than a preset threshold, and manually screening and confirming the marked structural damage image data, filtering out images unrelated to structural damage, obtaining and retaining the structural damage image with the best quality;

[0170] S22. Store the retained best-quality structural damage images in a separate folder to form the final three-level dataset.

[0171] It should be explained that the present invention also includes a comprehensive result output function. For image blocks classified by the secondary recognition sub-model, the results are marked on the original image and saved in the result folder. The marking color and text size are set according to preset rules. For example, cracks are marked with different shades of red, peelings are marked with different shades of green, and crushings are marked with blue.

[0172] Preferably, based on the deep learning classification algorithm and the three-level data set, a first-level recognition model and a second-level recognition sub-model are constructed; using the first-level recognition model to perform preliminary recognition and classification on the three-level data set to obtain damage-free image blocks and damaged image blocks, and inputting the damaged image blocks into the second-level recognition sub-model for secondary recognition and classification includes the following steps:

[0173] S31. Select the ResNet-50 architecture as the basic architecture, label the three-level dataset as damage-free and damage-affected, divide the best-quality structural damage image into several image blocks, and use each image block as the input of the first-level recognition model;

[0174] S32. Use ResNet-50 to process the input image blocks layer by layer to extract high-level structural features;

[0175] S34, use the fully connected layer to output the probability value of the two classifications, and convert the output of the first-level recognition model into a probability distribution through the softmax activation function;

[0176] S35. Use the binary weighted cross entropy loss function to measure the difference between the output of the first-level recognition model and the true label, and update the weight of the first-level recognition model through back propagation and gradient descent method;

[0177] S36, performing binary classification on all image blocks in the third-level data set based on the first-level recognition model to screen out damage-free image blocks and damaged image blocks;

[0178] S37. Based on the damaged image blocks, construct several parallel secondary recognition sub-models, and classify different types of structural damage into several damage levels through each secondary recognition sub-model.

[0179] It should be explained that the first-level recognition model and each second-level recognition sub-model all use the ResNet-50 model as the base model. Its core feature is its residual connection. These connections allow the network input to be directly added to the subsequent layers. This helps to solve the gradient vanishing and gradient exploding problems in deep networks, allowing the network to be trained more deeply. Specifically, it includes the following:

[0180] In the first layer, the processed image is input into the ResNet-50 model. Through the model's deep network structure, image features are extracted layer by layer. At the end of the network, a fully connected layer and a softmax activation function are used for classification prediction. The loss is calculated (using the cross-entropy loss function), and the model weights are updated through backpropagation and gradient descent. When the input image passes through each layer of ResNet-50, it is converted into a higher-level feature representation until a final classification result is obtained as to whether it is damaged or not.

[0181] Specifically, the features of the image are extracted layer by layer through multiple convolutional layers. Its main feature is the residual block, which allows the input to skip several layers and pass directly to the subsequent layers, thereby effectively solving the gradient vanishing problem in deep networks.

[0182] Preferably, ResNet-50 is used to process the input image blocks layer by layer to extract high-level structural features including:

[0183] The input image block passes through the convolution layer, pooling layer and residual block of ResNet-50 to extract low-level features layer by layer and gradually aggregate them into high-level structural features.

[0184] Preferably, the formula of the binary weighted cross entropy loss function is:

[0185]

[0186] Where Loss(1) represents the loss function value of the first-level recognition model;

[0187] N represents the total number of input image blocks for one training;

[0188] y i Represented as the true label of the i-th image block;

[0189] p i It is represented as the predicted category distribution probability of the first-level recognition model of the i-th image block.

[0190] Preferably, constructing a plurality of parallel secondary recognition sub-models based on damaged image blocks, and classifying different types of structural damage into a plurality of damage levels by each secondary recognition sub-model comprises the following steps:

[0191] S371. Based on the damaged image block, construct several parallel secondary recognition sub-models, and select a ResNet-50 architecture for each secondary recognition sub-model;

[0192] S372. Classify and label different types of structural damage, and divide structural damage into several damage levels.

[0193] Preferably, feature extraction and classification are performed on the results of the secondary recognition and classification based on a deep convolutional neural network, the category with the largest classification probability value is used as the damage type of the damaged image block, and the damaged image block is labeled; the damage-free image block is spliced ​​with the labeled damaged image block to obtain the final recognition result, which includes the following steps:

[0194] S41, based on the convolution layer of the deep convolutional neural network, a sliding window scan is performed on the input damaged image block, and a number of filters are used to extract the local features of the damaged image block;

[0195] S42, downsampling the feature damaged image block using a maximum pooling layer to reduce the size of the feature damaged image block;

[0196] S43. After multi-layer convolution and pooling, the extracted high-level feature damaged image is flattened into a one-dimensional vector and passed to the fully connected layer. The fully connected layer maps the feature damaged image to the final category space. Based on the Softmax activation function, the output value of the secondary recognition sub-model is converted into the probability distribution of each damage category.

[0197] S44. Use the multi-class weighted cross entropy loss function to measure the difference between the output of the first-level recognition model and the true label, and update the weight of the first-level recognition model through back propagation and gradient descent method;

[0198] S45. Based on the Softmax layer, the probability distribution of each type of damage for each damaged image block is output, the category with the highest probability value is selected as the final damage type of the damaged image block, and corresponding labeling is performed;

[0199] S46: splicing the intact image block with the marked damaged image block to form a result image consistent with the specifications of the original image as the final recognition result.

[0200] It should be explained that the working principle of deep convolutional neural network feature extraction and classification is as follows:

[0201] 1) Convolutional Layer

[0202] Feature Extraction: The convolutional layer is the core layer of a CNN, used to extract features from the input image. The convolution operation uses multiple filters to perform a sliding window scan of the image, extracting different features such as edges and textures. Each filter can capture different local features.

[0203] 2) Activation Function

[0204] After the convolutional layer, a nonlinear activation function (ReLU) is usually applied to introduce nonlinearity, allowing the model to learn more complex features. ReLU sets all negative values ​​to 0 and leaves positive values ​​unchanged, ensuring that the model can effectively distinguish complex patterns.

[0205] 3) Pooling Layer

[0206] Downsampling: Pooling layers are used to reduce the size of feature maps. Max pooling is typically used to reduce the feature dimension by taking the maximum value in a local region while retaining important features. Pooling can reduce computational effort and prevent overfitting.

[0207] 4) Fully Connected Layer (FC Layer)

[0208] Classification: After multiple layers of convolution and pooling, the extracted high-level features are flattened into one-dimensional vectors and passed to the fully connected layer, which maps these features to the final category space.

[0209] 5) Softmax layer

[0210] Probability output: After the fully connected layer, the Softmax activation function is typically used to convert the model's output values ​​into a probability distribution. Softmax converts the output value of each category into a probability value between [0, 1], where the sum of all probability values ​​for all categories is 1. The category with the highest probability value is the model's prediction.

[0211] Specific implementation steps include:

[0212] Step 1: Input data preparation: Preprocess the image into a format acceptable to CNN, including resizing, normalizing, and converting it into a tensor.

[0213] Step 2: Construct a deep convolutional neural network (CNN): This is the first and second level sub-network constructed by the present invention.

[0214] It's important to clarify that the primary network only identifies the presence of damage, but doesn't determine the type or severity of damage. It's the first layer of the network that receives the image under test. The secondary network, which matches and identifies different damage categories and their corresponding multiple levels, only feeds the portion of the image under test that has been identified as damaged by the primary network. The secondary network consists of multiple damage category recognition networks arranged in parallel.

[0215] Step 3. Define the loss function and optimizer: Select an appropriate loss function (cross entropy loss function) and optimizer.

[0216] Step 4: Train the model: Through multiple rounds of iteration, the model will be continuously optimized through forward propagation, loss calculation, backpropagation and weight update.

[0217] Step 5: Model Evaluation and Prediction: After training is complete, the model is used to perform classification predictions on new image patches. The model outputs the probability of belonging to each category, and the category with the highest probability is selected as the prediction result.

[0218] Step 6: Probability Interpretation and Result Output: The output of the Softmax layer is a probability distribution that explains the probability value of each category. This output can help us understand the model's judgment results for each category.

[0219] Preferably, the formula of the multi-class weighted cross entropy loss function is:

[0220]

[0221] Where Loss(2) represents the loss function value of each secondary recognition sub-model;

[0222] N represents the total number of input image blocks for one training;

[0223] M represents the total number of categories;

[0224] y i，c Represented as the true label of the i-th image block on the c-th category;

[0225] It should be explained that if the i-th image block belongs to category c, then y i，c =1, otherwise y i，c =0.

[0226] p i，c It is represented as the category distribution probability predicted by the secondary recognition sub-model of the i-th image block on the c-th category.

[0227] According to another embodiment of the present invention, Figure 2 As shown, a structural damage classification identification system based on a deep learning classification algorithm is also provided. The structural damage classification identification system includes:

[0228] Data acquisition module 1, used to acquire a structural damage image data set as a primary data set, and preprocess and enhance the structural damage image data to form a secondary data set;

[0229] Data screening module 2, used to perform preliminary screening on the secondary data set using a preset screening technology, and retain the best quality structural damage images as the tertiary data set;

[0230] Data classification module 3 is used to construct a primary recognition model and a secondary recognition sub-model using a deep learning classification algorithm and the three-level dataset; based on the primary recognition model, preliminary recognition and classification are performed on the three-level dataset to obtain undamaged image blocks and damaged image blocks, and the damaged image blocks are input into the secondary recognition sub-model for secondary recognition and classification;

[0231] The result recognition module 4 is used to extract features and classify the results of the secondary recognition and classification based on a deep convolutional neural network, taking the category with the largest classification probability value as the damage type of the damaged image block and marking the damaged image block; splicing the undamaged image block with the marked damaged image block to obtain the final recognition result.

[0232] In summary, with the help of the above technical solution of the present invention, the present invention realizes efficient and accurate detection of structural surface damage by constructing a first-level recognition model and multiple second-level recognition sub-models. It has the advantages of gradually refined detection, modular design, optimization of computing resources, and low misjudgment rate. It can greatly improve the accuracy and robustness of detection while ensuring the overall detection efficiency, and realize efficient identification of various damage types, locations and degrees of engineering structures, thereby improving the accuracy and efficiency of detection. The present invention adopts hierarchical detection and recognition. First, the first-level recognition model quickly screens out areas where damage may exist, reducing interference from irrelevant areas; on this basis, multiple parallel second-level recognition sub-models further refine the classification, thereby improving the recognition accuracy of different damage types and levels. Such a hierarchical structure significantly improves the overall detection accuracy, especially in complex and diverse damage scenarios, and can more accurately distinguish subtle damage types. By decomposing the detection process into two stages and combining efficient screening with fine classification, the present invention can efficiently and accurately identify various types of structural damage, and has strong scalability and adaptability. It not only improves the accuracy and efficiency of detection, but also screens out damaged areas through simple and fast binary classification, and then conducts more detailed detection. It can significantly reduce the computational burden and optimize resource utilization, which will significantly improve the intelligence and automation level of structural health monitoring and provide important technical support for the safety management of engineering structures.

[0233] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A structural damage classification identification method based on a deep learning classification algorithm is characterized by: The structural damage classification identification method includes the following steps: S1. Acquire a structural damage image dataset as a primary dataset, and preprocess and enhance the structural damage image data in the primary dataset to form a secondary dataset; S2. Use the screening technology integrated with Laplace transform algorithm to preliminarily screen the secondary data set and retain the best quality structural damage images as the tertiary data set; S3. Based on the deep learning classification algorithm and the three-level dataset, a primary recognition model and a secondary recognition sub-model are constructed; the primary recognition model is used to perform preliminary recognition and classification on the three-level dataset to obtain undamaged image blocks and damaged image blocks, and the damaged image blocks are input into the secondary recognition sub-model for secondary recognition and classification; S4. Feature extraction and classification are performed on the results of the secondary recognition and classification based on a deep convolutional neural network. The category with the largest classification probability value is used as the damage type of the damaged image block, and the damaged image block is labeled. The undamaged image block is spliced ​​with the labeled damaged image block to obtain the final recognition result. The method of constructing a first-level recognition model and a second-level recognition sub-model based on a deep learning classification algorithm and a three-level data set; performing preliminary recognition and classification on the three-level data set using the first-level recognition model to obtain damage-free image blocks and damaged image blocks, and inputting the damaged image blocks into the second-level recognition sub-model for secondary recognition and classification includes the following steps: S31. Select the ResNet-50 architecture as the basic architecture, label the three-level dataset as damage-free and damage-affected, divide the best-quality structural damage image into several image blocks, and use each image block as the input of the first-level recognition model; S32. Use ResNet-50 to process the input image blocks layer by layer to extract high-level structural features; S34, use the fully connected layer to output the probability value of the two classifications, and convert the output of the first-level recognition model into a probability distribution through the softmax activation function; S35. Use the binary weighted cross entropy loss function to measure the difference between the output of the first-level recognition model and the true label, and update the weight of the first-level recognition model through back propagation and gradient descent method; S36, performing binary classification on all image blocks in the third-level data set based on the first-level recognition model to screen out damage-free image blocks and damaged image blocks; S37. Based on the damaged image blocks, construct several parallel secondary recognition sub-models, and classify different types of structural damage into several damage levels through each secondary recognition sub-model.

2. The structural damage classification identification method based on deep learning classification algorithm according to claim 1 is characterized in that: The step of obtaining a structural damage image dataset as a primary dataset and preprocessing and enhancing the structural damage image data in the primary dataset to form a secondary dataset comprises the following steps: S11. Obtain a structural damage image dataset as a primary dataset and set up a generative adversarial network; Among them, the generative adversarial network includes a generator and a discriminator; S12, inputting the primary dataset as a training set into the generative adversarial network for training, using the generator to generate structural damage image data samples, which are then distinguished by the discriminator; S13. After the training is completed, use the generator to generate new structural damage image data samples, input random noise, add the new structural damage image data samples to the original primary dataset, and adjust the labels to form a secondary dataset.

3. The structural damage classification identification method based on deep learning classification algorithm according to claim 2 is characterized in that: The method of using the screening technology integrated with the Laplace transform algorithm to preliminarily screen the secondary data set and retaining the best quality structural damage images as the tertiary data set includes the following steps: S21. Using a Laplace transform algorithm to mark structural damage image data whose clarity is lower than a preset threshold, and manually screening and confirming the marked structural damage image data, filtering out images unrelated to structural damage, obtaining and retaining the structural damage image with the best quality; S22. Store the retained best-quality structural damage images in a separate folder to form the final three-level dataset.

4. The structural damage classification identification method based on deep learning classification algorithm according to claim 1 is characterized in that: The use of ResNet-50 to process the input image blocks layer by layer and extract high-level structural features includes: The input image block passes through the convolution layer, pooling layer and residual block of ResNet-50 to extract low-level features layer by layer and gradually aggregate them into high-level structural features.

5. The structural damage classification identification method based on deep learning classification algorithm according to claim 1 is characterized in that: The formula of the binary weighted cross entropy loss function is: Where, Loss (1) Expressed as the loss function value of the first-level recognition model; N It is represented as the total number of input image blocks for one training; y i Expressed as i The true label of the image patch; p i Expressed as i The first-level recognition model of each image patch predicts the category distribution probability.

6. The structural damage classification identification method based on deep learning classification algorithm according to claim 1 is characterized in that: The method of constructing a plurality of parallel secondary recognition sub-models based on damaged image blocks and classifying different types of structural damage into a plurality of damage levels by using each secondary recognition sub-model comprises the following steps: S371. Based on the damaged image block, construct several parallel secondary recognition sub-models, and select a ResNet-50 architecture for each secondary recognition sub-model; S372. Classify and label different types of structural damage, and divide structural damage into several damage levels.

7. The structural damage classification identification method based on deep learning classification algorithm according to claim 1 is characterized in that: The method of extracting features and classifying the results of the secondary recognition and classification based on a deep convolutional neural network, taking the category with the largest classification probability value as the damage type of the damaged image block, and marking the damaged image block; splicing the undamaged image block with the marked damaged image block to obtain the final recognition result includes the following steps: S41, based on the convolution layer of the deep convolutional neural network, a sliding window scan is performed on the input damaged image block, and a number of filters are used to extract the local features of the damaged image block; S42, downsampling the feature damaged image block using a maximum pooling layer to reduce the size of the feature damaged image block; S43. After multi-layer convolution and pooling, the extracted high-level feature damaged image is flattened into a one-dimensional vector and passed to the fully connected layer. The fully connected layer maps the feature damaged image to the final category space. Based on the Softmax activation function, the output value of the secondary recognition sub-model is converted into the probability distribution of each damage category. S44. Use the multi-class weighted cross entropy loss function to measure the difference between the output of the first-level recognition model and the true label, and update the weight of the first-level recognition model through back propagation and gradient descent method; S45. Based on the Softmax layer, the probability distribution of each type of damage for each damaged image block is output, the category with the highest probability value is selected as the final damage type of the damaged image block, and corresponding labeling is performed; S46: splicing the intact image block with the marked damaged image block to form a result image consistent with the specifications of the original image as the final recognition result.

8. The structural damage classification identification method based on deep learning classification algorithm according to claim 7 is characterized in that: The formula of the multi-classification weighted cross entropy loss function is: Where, Loss (2) Expressed as the loss function value of each secondary recognition sub-model; N It is represented as the total number of input image blocks for one training; M Expressed as the total number of categories; y i，c Expressed as i The image block in c The true labels of the categories; p i，c Expressed as i The image block in c The secondary recognition sub-model on categories predicts the category distribution probability.

9. A structural damage classification identification system based on a deep learning classification algorithm, used to implement the structural damage classification identification method based on a deep learning classification algorithm according to any one of claims 1 to 8, characterized in that: The structural damage classification and identification system includes: A data acquisition module is used to acquire a structural damage image data set as a primary data set, and preprocess and enhance the structural damage image data to form a secondary data set; A data screening module is used to perform preliminary screening on the secondary data set using a preset screening technology and retain the best quality structural damage images as the tertiary data set; The data classification module is used to build a primary recognition model and a secondary recognition sub-model using a deep learning classification algorithm and a three-level data set. The primary recognition model is used to perform preliminary recognition and classification on the three-level data set to obtain intact image blocks and damaged image blocks. The damaged image blocks are then input into the secondary recognition sub-model for secondary recognition and classification. The result recognition module is used to extract features and classify the results of secondary recognition and classification based on a deep convolutional neural network. The category with the largest classification probability value is used as the damage type of the damaged image block, and the damaged image block is labeled; the damage-free image block is spliced ​​with the labeled damaged image block to obtain the final recognition result.

Images